Incident Report: 20191104

Incident Report
We DDoS'd ourselves
Summary
Severity	Moderate
Impact	Medium (Manual uploaded non-functional, one canceled show as a result)
Event Start	04/11/2019 09:00
Event End	05/11/2019 05:00
Recurrence Mitigation	Stop the self-DDoS, restart samba, cry
Contacts
Recovery Leader	Marks Polakovs (MP)
Other Attendees	Matthew Stratford (MS), Isaac Lowe (IL), Anthony Williams (AW)

Summary

A combination of a scheduled offsite backup job being run irregularly and some strange behaviour by the studio PCs caused urybackup0 (specifically the music store) to have severely degraded performance. This in turn meant that music uploads to the central library, which normally take less than a minute, would take far longer, sometimes reaching 10 minutes and timing out. One show had to be canceled as a result because the host was not able to get all his music onto the system in time.

Resolution

The guilty Windows PCs were restarted, and the Windows 10 Photos app (was/will be?) uninstalled to prevent this from reoccuring.

Analysis

This performance degradation is not new - looking at xymon graphs (specifically, CPU rolling average and I/O on urybackup0) suggests that the issue has been going on for many months. Anecdotally, presenters have reported that music upload performance has been slow. This could also be related to the jukebox-more-like-jokebox issues, as it may have trouble fetching song files over NFS.

What made it particularly noticeable this time was the ZFS backup job, which was the straw that broke the camel's back. AW speculates this could have made the RAID controllers I/O bound and unable to service requests in time.

What Went Well

• Effective collaboration by the incident team, especially considering MP didn't really know what he was doing around backup0

What Did Not Go Well

• It took a long time to get the incident declared, by which point it had forced one show to cancel
• We had limited insight into what was going on with samba/zfs, limited to htop and tcpdumps, especially since...
• ... the current comp team aren't experienced with samba/zfs/backup0 and it took some time to find the people who knew what they were doing
• Red herrings abound (FFMPEG, RAID issues, hammering, the backup job - any or all of these could have been a red herring or a causatory factor)
• An unrelated issue with the MyRadio email system distracted the computing team

How We Got Lucky

• We found the people who knew what they were doing, and with their help brought it under control

Next Steps

Short Term

1. Remind presenters that computing are not infallible, and they should have all their music uploaded well in advance

• MP - done
• If the incident reoccurs, may be worthwhile reminding them about aux playout and manual tracklisting

2. Create a process of change management / logging, to ensure everyone is on the same page about system upgrades/changes - we had no knowledge about the changes to the backup script until IL informed us on IRC

3. Stop Windows 10 from being so stupid

4. Review regularly scheduled jobs and ensure as few as possible run during the daytime

Long Term

1. Upgrade Samba on urybackup0 to 4.10/4.11 - 4.6 is end-of-life (and has 10 CVEs)

• May need to wait until after term-time, as this is a risky process

2. Potentially move the Music Store and pool1 to separate RAID controllers in urybackup0

3. Or just take it behind the shed and put a bullet in its head - per AW, the chassis and RAID controllers are old

4. Review documentation and have URY computing alumni distill as much knowledge as possible

Timeline

(all times GMT)

9:00 - IL, after deploying a new backup script on urybackup0, notices a bug with it, fixes it, and runs it manually. It does a full backup, not an incremental one, partly to remediate said bug. It starts copying the entirety of the music store from backup0 to moyles (in IT Services TFTA data-center on East), which starts hammering the ZFS pools on backup0.

20:48 - A member reports in Slack that they are unable to upload music. With just over an hour to go to their show, and 12 songs left to upload, they are forced to cancel.

21:32 - MP alerts #computing in Slack to a serious incident. INCIDENT BEGINS.

21:35 - MS reports that he had done some package updates on dolby(?) over the weekend, including ffmpeg, which may be causing stuck uploads. (This is eventually shown to be a red herring, but is important to include in this report.)

Between these entries, the computing team are sidetracked by an unrelated issue with MyRadio emails

22:30 - MS reports that ffmpeg conversions are working manually.

22:36 - MP adds debug logging to the MyRadio upload controller, showing that it is hanging on a call to move_uploaded_file, a built-in PHP function. This leads him to point the finger to the music store.

22:55 - AW logs on to backup0 to troubleshoot the pools, sees the degraded performance, but notes that zpool reports it is running fine. AW notices the sluggish offsite backup job, notes that the RAID controllers are chugging as well. MP and AW decide to restart Samba, but have to wait until midnight as there is a show on air

23:13-23:25 - AW notices that a Samba log is generating 100KB/s, and a 5MB log size limit means that it is rotating the log every few seconds. He also notes that the ARC cache is very low and that swap is not enabled. AW notes that the incremental backup job... is not incremental... and has been sending the entire music dataset to moyles for over 12 hours. Around this time, a very frustrated MP goes for a run to let off steam.

23:35 - IL logs on and reports that the backup was intentionally not an incremental one. However, he notes that it isn't pulling enough data to be the sole cause of the slowdown.

23:41 - MP gets lost in Alcuin.

23:49 - MS and IL report that the Windows 10 photos app is seemingly hammering backup0's shares

00:05 - MP returns and stops samba. Performance on backup0 immediately jumps tenfold. However, MP screws up and doesn't realise that samba also handles his credentials (from the Windows domain).

00:10 - MS logs on with his local account and restarts samba. Performance degrades immediately. MP notes that Xymon does not monitor samba on backup0 at all.

00:14 - AW notes that the picture hammering noted earlier is likely one of the studio PCs.

00:38-00:43 - MP remotes into wogan and restarts all the studio presenter and guest PCs. At 00:56 he reports that performance has not, in fact, improved substantially.

01:06 - MS tests song uploading and notes that, while it is still slow, it has improved dramatically. INCIDENT ENDS

01:26 - MP and the gang head to bed.

Around 05:00 - the backup script finishes and network and I/O performance returns to normal.